Checkpoint Enum Switch 10X (Soze)

The usual way to swap checkpoints mid-workflow is a pile of loader nodes feeding a switch, which means every checkpoint you're "not using" is still sitting in VRAM. Checkpoint Enum Switch 10X inverts that: ten checkpoint slots, one string key, and only the matching model is ever loaded. The rest are just names in dropdowns - they never touch your GPU.

It's the checkpoint variant of Soze's enum-switch family (there are siblings for LoRAs, diffusion models and upscale models, all sharing the same matching logic and the same 10X shape). If you've ever wanted a CSV column or a text input to decide which checkpoint runs, this is the node that makes it painless.

How it works

You get enum_value (the key) and ten (compare_text_N, model_N) pairs. The node compares the key against each label - case-insensitive and whitespace-trimmed by default, first match wins - and loads only the checkpoint in the winning slot. default_model loads when nothing matches. Blank labels are skipped, so unused slots can't accidentally fire.

Two helpers make it usable rather than tedious:

• A "Copy Names to Compare Text" button (added by the pack's JS) fills every compare_text_N with the base filename of the model you picked in that slot, and copies a JSON array of those names to your clipboard. Paste that array into an enum_value source elsewhere in the graph and the whole lookup works.
• allow_none: flip it on and an enum_value of "none" - the first item in that copied array - skips loading entirely. The loader outputs nothing for that run. Handy when one row of your CSV means "no checkpoint change."

Outputs: MODEL, CLIP, VAE (the classic triple, ready for a sampler and CLIP text encode), plus model_name (which checkpoint actually loaded - wire this into your filename or a text log), matched_index, and status.

When it's worth it

Drive it from the pack's CSV Reader or CSV Random Reader: a prompt file where every row names its own checkpoint, and this node routes the right one in. Or use it as a tidy A/B test rig - enum_value flips between two labels, only one model in VRAM at a time, and model_name tells you which one produced the output. If you're mainly hand-selecting in the UI and never change mid-run, the built-in CheckpointLoaderSimple is honestly fine - this earns its keep the moment a string decides the model.

Installing it

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SozeInc/ComfyUI_Soze.git
pip install -r ComfyUI_Soze/requirements.txt

Or via ComfyUI Manager ("ComfyUI_Soze") and restart. No extra model files - it loads whatever's already in your models/checkpoints folder. If your dropdowns look empty, that's a checkpoint-folder problem, not a node problem: drop .safetensors files into models/checkpoints and refresh.

One trap: the model dropdowns live in the optional section, so if you collapse the node you may not see them. Leave it expanded while you're wiring the ten slots, and remember blank model dropdowns mean "unused."